Object Detection Based On Deep Learning And Application In Understanding Intelligent Vehicle Driving Environments

This paper focuses on solving two classical problems of object detection task in intelligent vehicle driving environment understanding:accuracy and speed of detection.In the driving process of the intelligent vehicles,it is essential to obtain the real time environmental information to judge the traveling direction and the speed.Therefore,the accuracy of detection determines the ability of the intelligent vehicles to perceive the surrounding environment.Meanwhile,the real-time feedback of detection results is needed by the detection network used in the driving process of the intelligent vehicles,so the high speed of the network is also required.Based on specific improvements on the basic detection framework,this paper establishes a targeted detection framework for understanding the driving environment of intelligent vehicle,and proposes a forced knowledge distillation algorithm for model compression and acceleration to achieve a fast and good detection model for traffic data sets,which meets the needs of intelligent vehicles for understanding the driving environment.The main research work of this paper is as follows:Firstly,this paper chooses the open traffic dataset KITTI as the research dataset and Faster RCNN as the basic detection framework.This paper improves the framework from the aspects of data processing,framework structure and training methods,and establishes a detection framework for understanding the driving environment of intelligent vehicles based on the basic framework Faster RCNN.In the aspect of data processing,data enhancement processing is introduced in my work;in the framework part,this paper adjusts anchor ratio,backbone network,and classification loss function for data set;add Hard Negative Mining method to deal with sample imbalance in model training;also use Soft-NMS algorithm to increase model's recall rate.In this paper,all the above improvements are fully verified on the KITTI dataset,and the targeted detection model is gradually optimized.The experimental results show that the targeted framework for the understanding of the driving environment of the intelligent vehicles has a powerful detection capability,which is 22.6%higher than that of the general framework.Secondly,this paper proposes a framework of forced knowledge distillation.The idea of knowledge distillation has achieved great success in classification tasks.This paper innovatively applies ideas to more complex target detection tasks,and proposes an end-to-end trainable knowledge-based distillation framework.At the same time,this paper compares various knowledge distillation methods and loss functions,and makes experiments to determine which combination is more suitable for detection tasks.Through experimental verification,the knowledge distillation method based on attention map and the normalized loss function can get the best results,and the result also verify that the forced knowledge distillation framework can greatly improve the network running speed without greatly reducing the accuracy,which can be effectively applied in the understanding of the driving environment of the intelligent vehicles.Finally,considering the need for fast and good detection results for intelligent vehicles driving,this paper merges the contents of the first two parts,establishing a targeted forced knowledge distillation framework.This paper does a number of experiments on KITTI dataset and makes a comparative analysis.The results turn out that the model trained in this paper has the ability to diminish the accuracy slightly while embracing large acceleration of detection,and even in some cases,it has a slight increase of accuracy.The results satisfy the high demands of detection performance for driving the intelligent vehicles,which makes the research closer to practical applications.